Agent Memory Hall — AI skill for Claude Code

Agent Memory Hall

**Your AI coding agent forgets everything the moment you close the session.** AMH fixes that — with a handoff protocol, not another vector database.

[](https://www.apache.org/licenses/LICENSE-2.0) [](https://modelcontextprotocol.io)

---

The Problem

Every developer who uses AI coding agents knows this ritual:

1. Work with Claude Code / Cursor / Copilot for two hours
2. Close the session
3. Open a new session — the agent has no idea what you were doing
4. Spend 10 minutes re-explaining the context
5. Repeat

Worse: the agent said *"I'll fix the auth bug next session."* It didn't. Because it forgot it promised.

**Existing memory tools solve storage** ("where do I put memories?"). AMH solves **continuity** ("did the agent actually follow through?").

---

What AMH Does

Three slash commands. One local SQLite file. Three minutes to set up.

/start    → reads last session's state, shows "here's where you left off"
/save     → checkpoint mid-session (after a milestone)
/wrap-up  → structured close: what was done, what's pending, what's next

Plus **Baton** — the part no other tool does:

Baton tracks:
  - open_loops:      "said I'd do X — haven't done it yet"
  - follow_ups:      "deferred 3 times, first seen June 10"
  - active_decisions: "chose A over B because..."
  - patterns:        "keeps making the same mistake"

When you `/start` the next session, Baton tells your agent: *"You promised to fix the auth bug. You haven't. It's been deferred twice."*

---

Quick Start (3 minutes)

Option A: SQLite Skills (zero dependencies)

**macOS:**

mkdir -p ~/.claude/skills/{session-start,session-save,session-wrap-up} ~/.claude/job-memo
python3 -c "
import sqlite3, os
db = os.path.expanduser('~/.claude/sessions.db')
conn = sqlite3.connect(db)
conn.execute('''CREATE TABLE IF NOT EXI